Is AI bad for the environment? Energy, emissions and the real verdict in 2026

Every time you ask a chatbot to rewrite an email, a chip in a distant building draws a little electricity, warms up, and needs cooling. So, is AI bad for the environment? A single prompt has a tiny footprint, roughly the energy of running a light bulb for a minute or two. But AI as an industry is growing so fast that it is pushing up electricity demand, keeping gas plants busy and turning big tech's climate pledges into uphill climbs, so its overall environmental impact is real and rising.

Both halves of that answer are true, and most arguments online come from people holding only one of them. In this explainer we look at where AI's footprint actually comes from (energy, emissions, hardware and, briefly, water), what the best published numbers say as of late October 2026, and where we land.

Short answer: is AI bad for the environment?

Per use, not very: Google says a median Gemini text prompt uses 0.24 watt-hours and emits 0.03 grams of CO2. In total, increasingly so: the International Energy Agency projects data centre electricity use will roughly double from 485 TWh in 2025 to about 950 TWh by 2030, with AI the main driver.

The environmental impact of AI, then, is less about your chat history and more about the scale of the building boom behind it, and about whether that boom runs on clean power or fossil fuels.

How it works: what does AI do to the environment?

An AI model is, in the end, a very large pile of arithmetic. Running that arithmetic takes specialised chips, usually GPUs, packed by the thousand into data centres. Those chips turn electricity into computation and heat. Everything else follows from that simple fact.

It helps to split AI's footprint into four parts:

• Electricity. Training a model is a one-off burst of computing. Using it, called inference, happens every time someone sends a prompt. The IEA's analysis, as summarised by Let's Data Science, puts inference at roughly 80 to 90% of AI computing today, so everyday use matters more than the famous training runs.
• Emissions. Electricity is only as clean as the grid it comes from. A data centre on hydro power in Norway and one on a gas turbine in Texas can do the same work with very different carbon costs.
• Hardware. Chips, memory and servers have to be mined, manufactured, shipped and eventually thrown away. This "embodied" footprint is easy to forget because it happens before the first prompt and after the last.
• Water. Data centres evaporate water to stay cool, and power plants use water too.

We will keep water short here, because our sibling explainer on how much water AI uses covers it in detail. The headline from that piece: per prompt, estimates range from about 0.26 mL (Google's on-site figure) to tens of millilitres once power-plant water is counted, while data centres worldwide consumed trillions of litres in 2025. The local question, whose river and whose aquifer, matters more than the global average.

An everyday analogy may help. Think of AI like air travel rather than like a single kettle. One kettle boil is trivial; so is one prompt. But nobody judges aviation by one passenger's seat. We judge it by how many flights there are, how fast that number is growing and what fuel the planes burn. AI is closer to that kind of problem.

Real examples: how AI is affecting the environment in 2026

The per-prompt numbers

The best per-use figures come from the companies themselves, which is a limitation in its own right. In August 2025 Google published a study saying the median Gemini Apps text prompt uses 0.24 Wh of energy and emits 0.03 gCO2e, which it compared to "watching TV for less than nine seconds", according to DatacenterDynamics. OpenAI's Sam Altman has written that "The average query uses about 0.34 watt-hours." Mistral, working with outside auditors, estimated 1.14 gCO2e for a 400-token response from its Le Chat assistant, a fuller life-cycle method that includes hardware manufacturing.

To put those in proportion, here is our own back-of-envelope comparison for someone who sends 50 text prompts every day for a year. The calculations are ours, based on the published per-prompt figures.

The UK comparison uses the roughly 7 tonnes per person figure cited by data scientist Hannah Ritchie, who concluded that regular text use of chatbots is a small part of a personal footprint. She is careful to add: "I am not saying that AI energy demand, on aggregate, is not a problem."

Two caveats matter. Google's study reported only a market-based carbon figure (which credits clean-energy purchases) and has not been peer reviewed, critics told DatacenterDynamics. And these are text prompts. Generating images or video, long reasoning modes and AI agents that run dozens of steps for one task can use many times more compute.

The totals

Zoom out and the picture changes. According to the IEA's April 2026 update, reported by Enlit, global data centre electricity demand grew 17% in 2025 while overall electricity demand grew 3%. Demand from AI-focused data centres grew 50%. The IEA projects total data centre use rising from 485 TWh to 950 TWh by 2030, about 3% of global electricity, with AI-focused use tripling to 465 TWh.

On emissions, the IEA's 2025 Energy and AI report says data centres account for around 180 Mt of indirect CO2 emissions, about 0.5% of global combustion emissions, rising to around 1% by 2035 in its base case and 1.4% in a faster "Lift-Off" case. A separate peer-reviewed study in the journal Patterns by Alex de Vries-Gao, published in December 2025, estimated AI systems alone could emit 32.6 to 79.7 million tonnes of CO2 a year, comparable to New York City's 52.2 million tonnes in 2023. He also noted how little companies disclose, calling it "highly problematic".

Big tech's climate goals under strain

The clearest real-world evidence comes from company reports. Google's 2026 environmental report showed its greenhouse gas emissions rose 18% in 2025, its largest annual jump on record, while electricity demand rose 37%, Axios reported. Google signed a record 12 GW of clean-energy deals but its share of carbon-free electricity stayed roughly flat.

Microsoft, which pledged in 2020 to be carbon negative by 2030, reported emissions up 25.1% in a year, from 16.2 to 20.3 million tonnes, according to The Next Web, putting it roughly 58% above its pledge baseline. Microsoft says the figure would have been closer to 34 million tonnes without its efficiency work and clean-power deals.

The gas problem

Where the electricity comes from is the swing factor, and in the US a lot of it is coming from gas. AFP, in a report carried by Khaleej Times, said research firm Cleanview had identified about 60 planned data centres with 97 GW of on-site, "behind-the-meter" generation expected into the early 2030s. Under conservative assumptions, those plants "would easily top 200 million tons of carbon dioxide emissions a year", the report said, along with local air pollutants such as nitrogen oxides and fine particles. Elon Musk's xAI, which trucked mobile gas turbines to its Memphis data centre in 2024, accounts for most of the roughly two gigawatts already online. The IEA has also flagged projects with on-site gas generation, "largely in the US".

The hardware footprint

Chips have a cost before they are switched on. A September 2026 report by a coalition including Greenpeace and Friends of the Earth, covered by Euronews, found Nvidia's supply-chain emissions rose from about 1.3 Mt CO2e in 2020 to 10.7 Mt in 2026, based on Nvidia's own sustainability reports. Nvidia does not disclose emissions from its chips in use; the report estimated those at 4 to 21 Mt in 2025.

Then there is the end of the line. A study in Nature Computational Science, reported by MIT Technology Review, estimated generative AI could add 1.2 to 5 million metric tons of e-waste by 2030, against more than 60 million tonnes of e-waste produced globally each year. Because AI hardware is often replaced within two to five years, how long servers stay in service matters a great deal.

Benefits and limitations: is AI really bad for the environment?

A fair assessment has to weigh the other side, and there is a real other side.

First, efficiency is improving fast. The IEA says the power consumed per AI task is falling "by at least an order of magnitude annually". Google reported that the energy and carbon footprint of its median prompt fell 33-fold and 44-fold respectively over 12 months. Mistral's published life-cycle data is a step towards a common standard.

Second, the tech sector is the biggest buyer of clean power. The IEA's 2026 update says it signed around 40% of corporate renewable power purchase agreements in 2025, and nuclear small modular reactor deals with data centre operators grew from 25 GW to 45 GW.

Third, AI might cut more emissions than it causes. The IEA estimates that adopting existing AI applications across industry, transport and buildings could lead to 1,400 Mt of CO2 reductions in 2035, about three times data centre emissions in its high-growth case. That is a big number, and it is the strongest argument AI's defenders have.

Now the limitations. That 1,400 Mt is a possibility, not a forecast: the IEA itself says "there is currently no momentum" to ensure widespread adoption, and that rebound effects (such as people switching from buses to self-driving cars) could cancel the gains. Efficiency per prompt does not cap total use, and history suggests cheaper computing tends to mean more computing. Clean-power contracts do not always mean clean electrons at the hour and place a data centre runs, which is why Google's emissions rose despite record deals. And the new gas plants being built today will run for decades.

So where do we land? We think the honest verdict is this: AI is not uniquely dirty, and your own text prompts are a small part of your footprint. But AI is currently making the climate problem harder, not easier, because demand is growing faster than clean supply. Whether it becomes a net positive depends on choices still being made about power sources, hardware lifespans and disclosure.

What it means for you: practical takeaways

If you came here feeling guilty about using a chatbot, the evidence suggests you can relax a little. Skipping a few prompts will not move the needle much compared with how you travel, heat your home or what you eat. That is not a licence for waste, though, and a few habits help both the planet and your productivity:

• Prompt well the first time. One clear request beats five rounds of regenerating. Our guide to writing better AI prompts shows how.
• Match the tool to the task. Long reasoning modes, image generation and especially video use far more compute than a quick text answer. If a regular search answers your question, use it. AI-powered search features such as Google's AI Mode add a model call to what used to be a simple lookup.
• Keep devices longer. Hardware is part of AI's footprint, and the same logic applies to your laptop and phone.

Where your influence is larger is collective. If a data centre is proposed near you, it is reasonable to ask where its power will come from, whether it plans on-site gas generation, and whether it will publish energy and water figures. If you buy AI services for a company, ask vendors for location-based emissions data, not just market-based claims.

What happens next

A few things look likely. Demand will keep climbing: the IEA expects data centre electricity to roughly double by 2030, and it notes that big tech capital spending passed $400 billion in 2025 with a further rise expected in 2026. Pressure for disclosure will grow, since researchers like de Vries-Gao and Shaolei Ren keep pointing out that AI-specific figures are mostly unpublished. And physical limits, from grid connections to local opposition, will shape where data centres go.

What we don't know is just as important. We don't know whether efficiency gains will outpace usage growth, or merely slow it. We don't know how much of the new gas capacity will be replaced by nuclear and renewables as they come online. And we are still guessing about AI's upside, because the emissions it might save in grids, factories and transport depend on adoption that has barely started. Our view is that the question "is AI bad for the environment?" is really a question about energy policy wearing a tech costume. The software is getting leaner; the power behind it is the part to watch.

FAQ

Why is AI bad for the environment?

AI runs on energy-hungry chips in data centres, and demand is growing faster than clean power supply. That means more electricity from fossil fuels, including new gas plants built for data centres, plus emissions from chip manufacturing, growing e-waste and local water use. Per prompt the impact is small; the problem is the scale of the industry's growth.

How does AI affect the environment?

Mainly through electricity use and the emissions that come with it. The IEA projects data centre power demand will roughly double to about 950 TWh by 2030. AI also needs hardware that has its own manufacturing footprint and becomes e-waste, plus water for cooling. On the other side, AI tools could help cut emissions in energy, industry and transport.

What does AI do to the environment?

Each AI request makes chips draw power and produce heat, which data centres remove with cooling systems that often use water. If the power comes from gas or coal, that adds carbon dioxide and local air pollution. Building the chips and servers creates emissions and waste too. Individually these effects are tiny; across billions of requests they add up.

How is AI affecting the environment?

In 2026 the clearest sign is in company reports: Google's emissions rose 18% in 2025 and Microsoft's rose 25%, both driven largely by data centre growth. In the US, dozens of data centres are planning on-site gas generation. At the same time, power use per AI task is falling quickly, so the outcome depends heavily on clean energy supply.

What is the environmental impact of AI?

The IEA puts data centre emissions at around 180 Mt of CO2, about 0.5% of global combustion emissions, rising to around 1% or more by 2035. One peer-reviewed study estimated AI systems alone could emit 33 to 80 million tonnes a year, similar to New York City. Water use and e-waste add to that, and estimates vary widely.